Explaining Runoff and Flood Drivers in the Pearl River Basin Using an Attention-Enhanced CNN–BiLSTM Model

Spatiotemporal heterogeneity in water availability and the coupled effects of meteorological and land-surface processes continue to complicate the diagnosis of runoff dynamics and flood behavior. Here we investigate surface-runoff variability across the Pearl River Basin (PRB), China’s second-largest river basin, using an attention-augmented deep learning framework (CNN–BiLSTM–ATT). Because consistent daily discharge records are unavailable for all sub-catchments, a GLDAS-derived runoff reference series is adopted as the prediction target. The hybrid architecture combines convolutional feature extraction, bidirectional sequence learning, and multi-head attention to represent nonlinear spatiotemporal controls on runoff. To enhance interpretability, SHAP-based attribution is applied to quantify the relative contributions of precipitation, temperature, soil moisture, wind speed, evaporation, and vegetation activity (NDVI) to modeled runoff and flood responses. The model attains a mean NSE and R² of 0.74 across sub-catchments, and the attention module improves robustness by reducing errors in several locations. Attribution results consistently identify rainfall and upstream inflow as the dominant drivers of runoff generation, with stronger soil-moisture influence in the eastern PRB and comparatively greater roles of evaporation and temperature in the western PRB. For floods, rainfall and upstream inflow remain the leading controls, while soil moisture emerges as the secondary contributor. Peak-flow analysis further shows significant positive associations of flood peaks with rainfall and soil moisture, whereas temperature, evaporation, wind speed, and NDVI exhibit no systematic linkage to peak magnitude. These findings provide an interpretable, basin-wide perspective on runoff and flood controls in the PRB, supporting process understanding and sustainable water-resources management.